高级检索

    基于句法增强的细粒度情感三元组抽取方法

    Fine-Grained Sentiment Triplet Extraction Method Based on Syntactic Enhancement

    • 摘要: 属性级情感三元组抽取(aspect sentiment triplet extraction,ASTE)任务主要是从句子中检测出属性词及其对应的评价词和情感倾向,然而当抽取多词属性词和评价词时,无法准确地抽取出全部的单词;当面对重复的属性词和评价词时,以往的研究很难学习到"属性词-评价词"词对之间所有的关联关系. 为解决这些问题,提出了一种基于句法增强的多任务学习框架,来解决端到端的情感三元组抽取任务. 句子中的句法结构反映的是句法属性和依赖关联信息,这对抽取任务和情感分类任务有积极作用. 该方法是利用依存句法嵌入图卷积网络充分挖掘句法特征,并将其传递到属性词抽取、评价词抽取和情感分析这3个子任务中,实现了句法信息与多任务联合学习框架的融合. 在情感分析任务的4个英文数据集和1个中文数据集上对模型进行了评估,实验结果表明,提出的方法是有效的且明显优于其他的基线模型,同时对具体案例进行分析,证明该方法一定程度上解决了多词和重复词的问题.

       

      Abstract: The task of aspect sentiment triplet extraction (ASTE) mainly detects aspect terms and their corresponding opinion terms and sentiment polarities from sentences. However, when extracting multi-word aspect terms and opinion terms, it is impossible to accurately extract all words. The existence of repeated aspect terms and opinion terms makes it difficult for previous studies to capture all the correlations between aspect terms and opinion terms in word pairs. In response to these problems, we propose a framework based on syntactic enhanced multitasking learning to perform the task of end-to-end sentiment triplet extraction. The syntactic structure of a sentence reflects syntactic attributes and dependency or association information, therefore having a positive effect on the extraction task and the sentiment classification task. The proposed model utilizes dependency syntactic embedding graph convolutional network to fully mine syntactic features in sentences, and then transmits these features to 3 sub-tasks including aspect terms extraction, opinion terms extraction and sentiment analysis, thus realizing the effective fusion of syntactic information and multi-task joint learning framework. The model is evaluated in the sentiment analysis task upon 4 English datasets and 1 Chinese dataset. The experimental results show that the proposed model is effective and significantly better than other baseline models. At the same time, the results of specific case analysis prove that the method solves the problem of multiple words and repeated words to a certain extent.

       

    /

    返回文章
    返回